The Agentic Coding Playbook — Free Guide by voxspek

DEV LOOP · V1

GitHub

Issues · PRs · Code

Source of truth

Identify the problem

Observe, test, notice what's off. AI agents can help surface anomalies.

+ AI assist

Define scope & spec

Vision, acceptance criteria, priority. AI drafts specs from your direction.

+ AI assist

Agent plans & executes

Planner clusters issues into sprints. Executor ships PRs overnight.

Fully automated · nightly

CI & automated tests

Catches regressions. Does NOT catch contextual drift.

Review, test & curate

Merge, flag, follow up. Test the result. Capture observations.

voxspek: voice → issues in 3s

Human judgment

AI agents

Automated

voxspek

Why GitHub as hub — not just prompting agents

Context persists. Issues live forever. Chat prompts disappear when the session ends.
Sprints are coherent. The planner sees ALL issues and clusters related changes together.
Teams work. Collaborators file issues and review PRs with zero special setup.
Audit trail. Every observation traces from issue → sprint → PR → merged code.
Cross-agent. Issues work with Claude Code, Cursor, Codex, or any future agent.
Better output. Structured issues produce better agent PRs than freeform prompts.

Download the playbook (.md) Download Claude Code setup guide (.md) Read online

The Loop — 5 steps

Step	Who	What happens
1. Identify	Human + AI	Observe, test, surface issues. AI agents can help flag anomalies, but human eyes catch contextual drift.
2. Define	Human + AI	Set vision, scope, acceptance criteria. AI drafts specs from your direction — you own the judgment.
3. Plan & execute	AI agent	Planner clusters issues into sprints. Executor ships PRs overnight. Code, tests, docs.
4. CI & tests	Automated	Catches functional regressions. Does NOT catch "correct but doesn't fit" — only you can.
5. Review & test	Human	Review PRs. Test the result. Curate. Capture observations for the next cycle.

voxspek lives at step 5 → GitHub: capturing your review and testing observations as structured issues that feed the next cycle.

What's inside the playbook

1

The architecture

GitHub as the message bus. Label schema. Issue metadata format.
2

Claude Code skills setup

CLAUDE.md template + agentic workflow skill file. Copy-paste ready.
3

The testing sweep

20 minutes, 15 observations. What to look for. Flow preservation.
4

Nightly routines

Planner (9 PM), executor (overnight), morning briefer (7 AM). Full configs.
5

The morning review

Tiered review: LOW auto-merge, MEDIUM/HIGH human review. ~30 minutes.
6

GitHub Actions glue

Auto-label issues from metadata. Auto-close sprint plans on merge.
7

Putting it all together

Day 1 checklist. Week 1 rhythm. What success looks like at day 30.

Where human judgment is irreplaceable

Your responsibility	Why agents can't do it
Vision & strategy	Agents don't know your market, users, or business model. They execute — they don't source intent.
Spec & roadmap	Agents draft specs from your direction. The direction — what to build and why — is yours.
Product curation	Does this feature fit? Does the flow make sense? Tests check function; you check context.
UI/UX commentary	Visual consistency, empty states, navigation flow. Agents don't perceive "this looks off."
Architectural review	Low-frequency, high-stakes branching decisions. Miss one and it costs six months.

Everything else — writing code, running tests, opening PRs, composing sprints — is the agent's job. The loop automates it. Your job is the five things above, done densely enough that the product doesn't drift.